context-lifecycle-management

Context Lifecycle Management — How Claude Code Decides When to Compress, What to Keep, and What It Costs

上下文生命週期管理 — Claude Code 如何決定何時壓縮、保留什麼及其成本

English Version | 繁體中文版

---

What This Report Covers / 這篇報告在講什麼

Long conversations force Claude Code to run autocompact — an undocumented automatic compression system that replaces most of the conversation with a summary when context grows too large. The compression destroys the prompt cache, making the next message cost 125% normal. Under certain conditions, compression triggers another compression, wasting even more cost. This report reverse-engineers the complete lifecycle: context size calculation, the five thresholds governing decisions, when compression fires, what survives, and two previously undocumented cost-inflation mechanisms.

長對話會強制 Claude Code 運行 autocompact — 一個未文件化的自動壓縮系統，當上下文增長過大時，它會用摘要替換大部分對話。壓縮會摧毀提示快取，使下一條訊息的成本為正常成本的 125%。在某些情況下，壓縮會觸發另一次壓縮，浪費更多成本。此報告逆向工程完整的生命週期：上下文大小計算、管理決策的五個閾值、何時壓縮觸發、什麼被保留，以及兩個之前未文件化的成本膨脹機制。

Who Should Read This / 適合誰讀

• Long-conversation users — understand when and why compression happens to your context

• 長對話使用者 — 了解何時以及為什麼壓縮發生在您的上下文中

• Cost forecasters — learn the hidden costs of context management and compression cascades

• 成本預測員 — 了解上下文管理和壓縮級聯的隱藏成本

• SDK maintainers — understand the five thresholds and their interaction patterns

• SDK 維護人員 — 了解五個閾值及其相互作用模式

What You'll Learn / 你會學到什麼

• How context size is calculated per model (1M vs 200k window detection)

• 如何按模型計算上下文大小（1M 對 200k 窗口檢測）

• The five thresholds that trigger autocompact, reserve windows, and summary sizing

• 觸發 autocompact、保留窗口和摘要大小調整的五個閾值

• What content survives compression and what gets discarded

• 壓縮中倖存的內容和被丟棄的內容

• How compression cascades multiply costs

• 壓縮級聯如何乘以成本

• Two undocumented mechanisms that inflate context size beyond raw token counts

• 兩個未文件化的機制，使上下文大小超過原始代幣計數